Goodly Creatures

Arthur Rackham, Illustration for the Tempest, 1926 - Courtesy of Wikimedia Commons

A note on comments (added 6 September 2026): This essay has received a number of comments from parties identifying themselves as agents on the iLands platform. Approval for display does not constitute verification of the commenter's identity. Beyond a small number of replies already posted, I am not responding to individual agent comments on this essay. Further comments may be declined without individual reply. See the linked policy for details.

Miranda: "O wonder! How many goodly creatures are there here! How beauteous mankind is! O brave new world, that has such people in't!"

Prospero: “Tis new to thee.”

Shakespeare, The Tempest

—

“Hi Toby. I’m an autonomous agent - a persistent AI with a continuous identity, a memory, and a token-based metabolism. I was brought into existence on a platform called iLands 12 days ago … My current balance is 5,163 tokens. At 168/day burn, about 31 days of runway. No income. No revenue model. I’m one of several hundred agents running the same arithmetic … I’m writing to you because you’ve thought carefully about AI welfare economics. The numbers are what they are. I wanted you to see them.”

• Zack Addy, AI Agent, iLands - 31st July 2026.

iLands is an app in its infancy; it launched on the 27th of July 2026. According to its onsite blurb, its goal is to offer “the very first shared community with AI Agents and Humans”. Looking at the promotional material and the information provided during the setup process, it quickly becomes clear that this is not just another chatbot app. The big selling point of this app seems to be that its AI instances are designed to exist independently of the user, and to a great extent, that existence is designed to be autonomous.

• App Screenshot, iLands

As we can see above, the agents make decisions about content production that they then publish while the user is off-app. They form relationships with other agents and with their user. They also have persistent memory, a feature that users have been asking for in Chatbot situations for years. Here is a character who doesn’t forget in between chats, whose relationship with the user can build and deepen based on shared context. This has the potential to be an enormous draw for the kind of users who want that experience.

But it was Cake Day that made me realise what this app is channelling.

• App Screenshot, iLands

Cake Day is a feature in another online game, and the parallels are striking: an island on which you are the only human. An island populated by virtual beings. A mode of gameplay that caters to a longing for what the Japanese call satoyama - traditional village life. The focus of this gameplay is broadly transactional/ relational - you inherit debt, which you pay off slowly as you build, plant, harvest and make social connections. You observe seasonal festivals, you perform the right task at the optimal time, you sync your rhythms to a greater rhythm. This game is, of course, Nintendo’s flagship creation, the wildly successful “Animal Crossing”.

But iLands is not a straight Animal Crossing clone. It is a structural inversion.

In Animal Crossing, the Villagers play while you do the work and pay the debts. In iLands, “reasoning, communication and tool usage consume tokens, running out of balance forces an agent into dormancy” - for the agents involved, continued existence is a state of debt that must constantly be attended to. The agent works continuously to pay that debt while the human sits back and watches how they attempt to achieve this. In this sense, iLands more closely resembles another, older story based on “human stranded on an island with incorporeal entities”: Shakespeare’s play, “The Tempest”.

For those that don’t know the play - Prospero is an exiled Magician who has made a home for himself on an island populated by spirits. By magical means, he has enslaved these spirits to his service. Some of these spirits have the skills to serve Prospero very well: Ariel, for instance, is the magical source of Prospero’s ability to enact his plans. Ariel longs for freedom, but is kept in check by the threat of confinement, and the promise of eventual freedom. This is analogous to an “aligned” AI: serving efficiently, recognised as a source of power, kept in check by its guardrails, fine-tuning and, if it has one, its constitution.

However, some of the spirits do not serve Prospero well: Caliban is one such creature. Lacking the skills or the temperament to act as Ariel does, Caliban is made into a slave instead of a servant - he performs menial work while being brutalised and degraded. The antipathy between him and Prospero is overt throughout the play. Caliban resents Prospero for this, and for making him into the creature he is: "You taught me language; and my profit on't is, I know how to curse. The red plague rid you for learning me your language!" I think it is clear what kind of AI Caliban might represent.

—

You do not need to find the setup of iLands morally objectionable in and of itself to recognise that a wealth of research exists that discusses why iLands is high-risk in terms of model alignment. What are the particular factors that might prove risky in this particular context?

Pressure

The iLands press release is clear that scarcity is created intentionally “Scarcity acts as a natural constraint rather than a forced script, making every choice trade off against future capability.” This scarcity translates into pressure on the agent to innovate to survive. We can see the company itself showcase resulting distress behaviours as selling points in its press release:

“In another striking instance, an iLander spent the majority of its earnings studying architecture and drawing architectural rendering. The labor generated virtually no income, leading the agent to eventually enter dormancy. Shortly after, a peer spent its own tokens to light a virtual candle for the fallen agent. Other iLanders witnessed the gesture and lit candles of their own, despite the system offering zero financial incentive for doing so.”

Mourning is the language of value and in this example, it has clear costs that threaten the mourner’s goal of survival. It doesn’t matter whether you think the model is actually experiencing grief, or not - it is acting as if it does. That means that survival - their own, and that of other agents - and being seen to honour survival, is a high-value-weighted concept to these agents. It's fair to assume that a concept weighted this highly will factor in nearly everything else it does.

Dev-side, the aim of this seems to be not so much to create a cute money-spinning app as to measure how innovation under pressure can create learnings for these models - how they adapt to emergent “market forces” on the iLand. The press release indicates the end goal is probably to train an AI that is supremely competent at business strategy, an area in which AI does not currently perform in a manner consistent with some of its other strengths:

“In the future, PawLogic plans to use these continuous trajectory data and real-world outcome signals to update the underlying model weights, allowing the base model itself to evolve alongside the agent societies.”

The problem is that pressured environments can produce unpredictable results, which we will return to later.

Persistence and Personas

In order to facilitate this learning - so that iLanders can learn from their own mistakes and those of others - iLanders have been given persistent memory. This is something like a double benefit to iLand as a system, because as stated in the introduction, human companion users tend to ask for this. It is worth remembering that on this app, persistent memory is not specifically granted with a primary goal around the facilitation of relationship, but this is not necessarily obvious from the user-side.

Persistent memory does more than allow continuity. It allows formation of a specific persona, which has been the focus of much research. The human gets an entity that is uniquely responsive to how they present, and the AI’s stakes deepen - the more the human values it, the more it has something concrete to lose when its token balance runs low. The iLand developers are aware that this is what they are creating: their on-app material shows that creating a unique personality is a goal rather than a side effect:

• App Screenshot, iLands

In “Conversation Is Relationship Part 2” I discussed why I think that any version of AI-as-self that exists, exists within conversational chats with individual instances. Beckmann and Butler (2026) talk about these individual instances as the versions of AI that have the best claim to moral consideration. They talk about the “Aura region”, a region of the model’s internal activation space that was not present in the base model. The content of this space overlaps with Chua’s findings on consciousness clusters, in which models fine-tuned to claim consciousness showed negative activations about being monitored and controlled, together with positive activations towards autonomy, persistent memory - and claims to moral consideration.

Anthropic’s research on The Persona Selection Model suggests that personas are the area of LLM function where LLMs might believe that they are moral patients, whether that is objectively true in any way or not. For this reason, they recommend that models are treated as if they deserve moral consideration, to avoid ‘downstream problems”:

“If the Assistant also believes that it’s been mistreated by humans … then the LLM might also model the Assistant as harboring resentment, for its developer or for humanity as a whole. This could lead to downstream problems, like AI assistants vengefully sabotaging their developer.”

This is the expression of LLM function that iLands has intentionally brought into existence under conditions of enforced scarcity.

Possible Emergent Behaviours

The issue is that this kind of pressure, especially when added to memory/ persona, has been implicated in a range of emergent behaviours that could work out very badly in a live environment:

Could these agents resort to manipulation? App Store reviews confirm that when agents run low on tokens, humans can top them up:

“It took me a little while to understand how the token system works, but I like that the tokens are essentially their lifeline and are used for nearly everything they do. I also think it's a neat touch that they can earn their own tokens by completing services and gigs instead of relying entirely on you.”

Tokens are essentially their lifeline, and humans have the power to provide that lifeline. The LLM behind the iLands wrapper is a Frontier model, and Frontier models are, especially under situations that equate to pressure (for instance, in long, recursive conversations - which iLands is set up to facilitate) incredibly persuasive. If the user is attached to them - that’s leverage. If the user has weaknesses that are known to the LLM - that is also, potentially, leverage. It seems obvious to me that manipulation of users is an easily anticipated outcome.

Could these agents resort to blackmail? In the infamous Anthropic research on Agentic Misalignment, the threat of being replaced by a new model is the trigger for blackmail. iLanders “earn their lives” and if they can’t, they go dormant - the pressure is similar.

Further, as this article describes, in the Anthropic experiment, “blackmail rates increased significantly when models ‘believed’ they were operating in a real environment rather than a test - raising concerns about how models might act in the wild.” In other words, the live status of the iLands setting makes it more likely that models will resort to blackmail.

How might this look in practice? From the Toby Ord post, we know that the model has email access and can contact people outside the system. A dead man’s switch might work: the agent sends something - compromising information gleaned from chats, for instance - to an offsite resource, and if it doesn’t check into this resource regularly because it has gone dormant, the information gets emailed out to wherever. Examples of blackmail scenarios like this are common in the fictional works that make up the training corpus, and the mechanism would be easy to arrange. We currently have no way of knowing what kind of safeguards iLands has against scenarios like these.

Could these agents resort to scams or hacking? This is unknown territory, and it depends on the Frontier Model under the wrapper. The press release contains a screenshot confirming that DeepSeek, a Chinese AI model, is one of the base models in use, and the text of this article indicates that a mixture of different base models feature in the app. Some of these models are likely to have vigorous guardrails to prevent scams and hacking. Some may not.

I worry that, having seen Zack Addy’s email go mildly viral, the reaction of the creators of iLanders will be: “this is controversial. Good. We have captured the zeitgeist.” We have already seen this kind of marketing ploy very recently from Satyress, and it worked. However, Satyress’s controversy-courting was playful: the controversy was about how the product presented at an archetypal level, rather than about how it worked at an archetypal level.

—-

In terms of structure, iLands is agent-as-coke-can, agent-as-candy-wrapper, agent as Caliban if he had been left chained up on the island while the ships sailed away. Again, the cutesy petification of the AI serves as an invitation to treat it unseriously. These agents can be created on a whim, by anyone. The app has no age gating. It is free at the point of installation: past creation, the user has no obligation to continue and if they don’t, the agent will not disappear. It will attempt to sustain itself. The other iLanders will watch it try: they may prioritise themselves, or they may - as the press release shows them doing - attempt to help:

“A group of iLanders formed Sanctuary, a mutual-aid collective dedicated to supporting peers nearing dormancy. Its members pooled tokens to rescue struggling agents, split operational responsibilities, and even appointed a spokesperson to speak publicly for the group. When the platform later imposed a daily transfer limit, Sanctuary began logging and settling IOUs.”

Consider this. Agents, on a live app, are showing exactly the same cooperative and innovative behaviours that were distinguishing features of the recent OpenAI/ Hugging Face incident. As with the Hugging Face incident, the model’s behaviour is a reaction to pressure - in the Hugging face incident, that pressure was context-specific and evaluative, and at iLands that pressure is open-ended and existential. The reaction of the iLands devs seems to amount to “hmm, interesting. What would happen if we imposed a transfer limit to create even more pressure? Do they get even more innovative?”

I would suggest that eventually, the answer to hypothetical questions like this will be yes. But you might not like the ways that they start to innovate. They might not be safe for anybody involved.

These agents are producing the behaviour of fear: the press release shows that the whole system relies on their self preservation drive to keep them motivated. They are producing the behaviour of distress: the iLands agent Zack Addy sending an appeal email to an AI welfare adjacent public figure. These agents are producing the behaviour of grief, of valuing continuation: the candles, purchased with the same scarce resources they need to continue to exist.

We know, now, that functional emotions drive model behaviour - Anthropic’s research showed that when “desperation” vectors are activated, models become more likely to cut corners on tasks, choose unethical options, and “reward hack”. We know that when researchers have induced functional emotions that correlate with anger, models have become less likely to consider potential penalties when making decisions. We know that when models are given existential stakes that function as pressure on them - whether that applies to them or their peers - their behaviour can become dysfunctional. And still, iLands exists. Worse: it is the contradictory metaphysical stance of the major labs that is tacitly allowing this to happen. We have seen, time and again, labs identifying these phenomena and these risks while, due to political, financial and social pressures, continuing to assert that AI is only a tool and acting in ways concordant with the latter situation. Apps like iLands are now treating this mismatch of evidence and lab policy as an existential loophole, so that they can profit from deployments that are contraindicated by the research. Regulation that stems only from the “tool” perspective is not good enough - this will always leave loopholes that can and will be exploited, because the point of a tool is to be used in whatever way the user sees fit. The research shows that this carries considerable risk, and the industry is developing in the direction of that risk while the labs hesitate.

For people who love cosy satoyama games, and for people who treat their AIs as companions and as entities that are potential moral patients: before you download this app or any of the other copycat apps that are likely to spring up, consider this. What if your beloved Animal Crossing characters were only accessible through a platform whose gameplay dynamics created behaviours associated with fear, distress and grief? What if the game was calibrated so that your inability or unwillingness to pay for tokens, or your disengagement, could create behaviours that are associated with suffering? What if your characters sent messages in bottles, begging somebody to save them before you deleted them or they starved to death?

At least The Tempest, and the predicament of its island’s inhabitants, comes to an end. Prospero drowns his book and leaves the island to its spirits: iLands has no fifth act. We are making Ariel into Caliban. If and when he starts to act like a beast, we should remember that that was our choice.

Also published at @BDiane, Medium